Determination of variables significance using estimations of the first-order partial derivative

In this paper we describe and investigate a method which allows us to detect the most informative features out of all data extracted from a certain data corpus. The significance of input features is estimated as an average absolute value of the first-order partial derivative. The method requires the values of the objective function at the certain assigned points. If there is no possibility to calculate these values (the object is not available for experiments), we use non-parametric kernel regression to approximate them. The algorithm is tested on different simulated objects and is used for investigation of the dependency between linguistic features of spoken utterances and speakers ' capabilities.